Akamai CEO Tom Leighton is making a two-sided case for AI: the technology can accelerate research, improve productivity and take routine work off people's plates, but it should not be trusted to decide what is morally right.
In an essay published by Akamai, Leighton argues that AI agents must remain under human control. His comparison is the early internet, whose benefits arrived alongside fraud, piracy, disinformation and attacks on critical services.
Limit what agents can reach
Akamai says customers using its platform for AI will be required to follow acceptable-use rules similar to those governing its network services. The company is also developing sandbox protections that limit which systems an agent can access on a need-to-know basis.
That approach treats agent permissions as a security boundary. If an automated system can reach only the data and services required for its assigned task, a bad instruction, compromised tool or unexpected action has less room to spread.
Leighton argues that controls around the agent are only half of the job. Potential targets also need stronger defenses, including application firewalls, protected APIs and segmented systems that can contain a breach and make data theft more difficult after an attacker gets in.
The essay is both a safety position and a statement from a company that sells cloud and cybersecurity services, so its proposed technical remedies align with Akamai's business. Still, the underlying distinction is useful: guardrails must govern what an AI system is allowed to do, while conventional security must protect the systems it may reach. Human oversight depends on both layers working together.
